Does the Pytorch Lightning Trainer use the validation data to optimize the models weights?

Question

I am currently working with Pytorch Forecasting, which heavily uses Pytorch Lightning. Here I am applying the Pytorch Lightning Trainer to train a Temporal Fusion Transformer Model, roughly following the outline of this example. My rough training code and model definition looks like this:

training = TimeSeriesDataSet(
    df_train[lambda x: x.time_idx <= training_cutoff],
    time_idx="time_idx",
    target="target",
    group_ids=["group"],
    max_prediction_length=90,
    min_encoder_length=365 // 2,
    max_encoder_length=365, 
    time_varying_unknown_reals=["target"], 
    time_varying_known_reals=["time_idx"]
)

validation = TimeSeriesDataSet.from_dataset(training, df_train, predict=True, stop_randomization=True)

# create dataloaders for model
batch_size = 4  
train_dataloader = training.to_dataloader(train=True, batch_size=batch_size, num_workers=0)
val_dataloader = validation.to_dataloader(train=False, batch_size=batch_size, num_workers=0)

tft = TemporalFusionTransformer.from_dataset(
    training,
    learning_rate=res.suggestion(),
    hidden_size=16,
    attention_head_size=1,
    dropout=0.1,
    hidden_continuous_size=8,
    output_size=7,  
    loss=QuantileLoss(),
    log_interval=10,  
    reduce_on_plateau_patience=4,
    time_varying_reals_encoder=["target"],
    time_varying_reals_decoder=["target"]
)

trainer = pl.Trainer(
    max_epochs=15,
    gpus=0,
    weights_summary="top",
    gradient_clip_val=0.1,
    limit_train_batches=30,
    callbacks=[lr_logger, early_stop_callback],
    logger=logger,
)

trainer.fit(
    tft,
    train_dataloader,
    val_dataloader
)

Now my question is, whether the validation data has any influence on the optimization of the model? I have been playing around with the max_prediction_length parameter and it seems to be the case that the model performs better when I set the validation time window to a larger time frame. Does the Pytorch Lightning Trainer use the validation data to optimize the model or am I missing something else?

Thanks a lot in advance!

Answer 1

Since PyTorch-forecasting is an abstraction built on PyTorch-lightning, we can refer to the documentation on the trainer abstraction of the latter framework (https://pytorch-lightning.readthedocs.io/en/stable/common/trainer.html).

# put model in train mode
model.train()
torch.set_grad_enabled(True)

losses = []
for batch in train_dataloader:
    # calls hooks like this one
    on_train_batch_start()

    # train step
    loss = training_step(batch)

    # clear gradients
    optimizer.zero_grad()

    # backward
    loss.backward()

    # update parameters
    optimizer.step()

    losses.append(loss)

In the example above, we can see that the trainer only computes the loss of batches in the train_dataloader and propagates the losses back. It means that the validation set is not used for the update of the model's weights.